GWM Worlds 2 vs Genie 3: which real-time world model to use

These are the two closest models among those compared on this site. Both generate a world you walk into rather than watch. Both stream 720p at 24 frames per second in real time. Both were built by labs with more compute than almost anyone else, and both refuse to give you a file at the end. The differences are narrow, and they are decisive. This is an independent comparison, not affiliated with Runway or Google DeepMind.

The specifications that are identical

Start with what does not separate them, because it is most of the headline numbers.

Genie 3 renders at 720p and 24 fps, and DeepMind states it stays consistent for several minutes as you move through it (DeepMind). GWM Worlds 2 renders at 720p and 24 fps as one continuous stream (Runway). Identical, and not by coincidence.

Both are bounded by the same constraint: a frame has to arrive before a human notices the delay. That budget, not model quality, sets the resolution. It is also why the offline models look so much better on paper, with Atlas generating at up to 1440p because nothing is waiting on it, as covered in Atlas vs Genie 3.

Neither exports 3D. Both give you frames and nothing else.

Audio: the first real difference

GWM Worlds 2 generates audio at 48,000 Hz alongside the video, from the same foundational audio-visual model (Runway on X). Footsteps, voices and ambience come out of the same process that makes the picture, rather than being added afterwards.

Genie 3 does not generate sound. Nothing in DeepMind’s material describes audio at all.

For research into agent behaviour this may not matter. For anything resembling entertainment it is close to decisive: a silent playable world is a tech demo, and an audible one is a scene. It is the single clearest reason to call GWM Worlds 2 the more advanced model on published capability.

Actions: the second real difference

Genie 3 gives you navigation plus what DeepMind calls promptable world events: you move through the world, and you can type a change, a shift in weather, a new object, mid-session. The vocabulary of movement is fixed; the vocabulary of events is open but operates on the scene rather than on its inhabitants.

GWM Worlds 2 opens both. Subjects and the scene can be directed to take arbitrary actions through text, addressed to any subject individually or to the world as a whole, alongside continuous camera motion. Runway states this is the first world model to generalise to arbitrary actions rather than a fixed set or navigation alone (Cristobal Valenzuela).

The practical difference is whether you are a visitor or a director. In Genie 3 you explore a place that changes when you ask. In GWM Worlds 2 you tell a character to climb the wall and the character climbs the wall. How the world is set up in the first place, through the genesis prompt, is covered in the GWM Worlds 2 guide.

Caveat worth stating plainly: the arbitrary-actions claim has been demonstrated by Runway and tested by nobody else. Generalisation claims are exactly the sort that degrade when strangers start prodding them.

Consistency: different mechanisms, different failure modes

Genie 3 maintains what DeepMind calls long-horizon memory during interaction. Turn away from a wall and back and the wall is still there, minutes later. DeepMind qualifies this as holding for several minutes, after which drift can appear.

GWM Worlds 2 conditions each autoregressive step on three things: global context, meaning the genesis prompt and first frame; the current frame’s inputs; and past generated frames held in a sliding window. Runway states there is no preset session length.

Those two statements are less far apart than they look. A sliding window is finite by definition, so GWM Worlds 2 has a memory horizon too; Runway simply has not published its size. The honest reading is that both models forget, both were demonstrated over minutes, and neither has been stress-tested in public over an hour.

Availability: where Genie 3 wins outright

Genie 3 reached the public as Project Genie for Google AI Ultra subscribers in the United States on January 29, 2026 (Wikipedia). If you have that subscription and you are in the US, you can be inside a Genie 3 world in a few minutes, at no marginal cost.

GWM Worlds 2 is a research preview. There is no self-serve access, no pricing, no subscription that includes it and no API; enquiries go through a business contact form. The full picture is in GWM Worlds 2 pricing and access.

Neither has a developer API, so neither is buildable. But one of them is at least usable, and that is worth more than a specification advantage to most people reading this.

Head to head

	GWM Worlds 2	Genie 3
Vendor	Runway	Google DeepMind
Announced	Sep 3, 2026	Aug 2025
Video	720p, 24 fps, real time	720p, 24 fps, real time
Audio	48,000 Hz	None
Actions	Arbitrary, addressed to subjects or scene	Navigation plus promptable world events
Session	No preset length	Consistent for several minutes
3D export	None	None
Access	Research preview, contact form	Google AI Ultra subscribers, US
API	None	None

Sources: Runway, DeepMind.

Which to choose

• You want to try a playable world model this week: Genie 3, if you are a US AI Ultra subscriber. It is the only one you can enter.
• You are evaluating what the category will look like in a year: GWM Worlds 2. Audio and open actions are where this is heading.
• You are building a product: neither. No API, no pricing, no terms. Look at the comparison hub for models you can actually call.
• You need assets for an engine: neither, again. Both are frame generators. Atlas and Marble export splats; see the game development guide.
• You are training embodied agents: Genie 3 today for access; GWM Worlds 2 later for action diversity, as discussed in the robotics real-to-sim guide.

What would change this comparison

• A Runway API. On published specifications, GWM Worlds 2 would go from unreachable to the strongest option among the playable models compared here.
• Audio in Genie 3. DeepMind closing that gap removes the clearest published advantage on the Runway side.
• A published memory bound from either lab. Whoever states how long a world actually stays coherent will set the benchmark everyone else has to answer.
• Independent testing. Neither model has been measured by a third party on a shared task. Every comparison above, including this one, rests on vendor claims.

What each lab says the model is for

Positioning is not marketing noise here; it predicts what each model will get good at next.

Runway frames GWM Worlds 2 as a foundation for interactive entertainment, virtual characters, robotics and embodied-agent simulation. Three of those four are about inhabited worlds rather than navigable ones, which explains why the model carries voices in the genesis prompt and why actions can be addressed to individual subjects. Runway is a company that sells to filmmakers and studios, and the model reflects that.

DeepMind frames Genie 3 as a step toward training embodied agents in simulation. The emphasis is on the environment as a training ground rather than as a stage: consistency over minutes, promptable events to vary conditions, and a research programme that runs through agent learning rather than content production.

The overlap is robotics and embodied agents, where both labs are pointing. The divergence is everything else: one lab is building somewhere for characters to perform, the other is building somewhere for agents to learn.

If you are choosing between them for a use case that is neither, expect the gap to widen rather than close. Audio and character direction are not features DeepMind appears to be pursuing, and long-horizon research-grade consistency is not obviously Runway’s priority either.

Why both landed on exactly 720p and 24 fps

The matching specifications are not a coincidence or a shared architecture, they are a shared constraint, and understanding it tells you what to expect from the next generation.

An interactive world model has a per-frame compute budget set by human perception. At 24 fps you have roughly 40 milliseconds to produce a frame before interaction starts to feel wrong, and that budget has to cover the full generation step. Everything else, resolution included, is what is left after paying for latency.

That is why the offline models look so far ahead on paper. Atlas generates at up to 1440p because nothing is waiting on it; it can spend seconds per frame. The comparison in Atlas vs GWM Worlds 2 is really a comparison between two different budgets rather than two different quality levels.

The practical forecast: interactive world models will reach 1080p when either the models get cheaper per step or the hardware gets faster, and both labs will arrive at roughly the same time, because they are solving the same equation. Resolution is therefore a poor axis on which to choose between them. Audio, action vocabulary and access are the axes that actually separate these two.

Bottom line

GWM Worlds 2 is the better model on paper and Genie 3 is the better model in practice, and that gap closes the moment Runway opens access. If you want to understand where interactive world models are going, read the GWM Worlds 2 material; if you want to spend an evening inside one, open Project Genie.

For the exportable side of the field, which neither of these models serves, see Atlas vs GWM Worlds 2. To hear when Atlas access widens, follow the changelog or its RSS feed.